Personality

Every AI system has a personality. Some of it comes from the core model (pretraining distribution, instruction tuning, RLHF), and some from the surrounding scaffolding (system prompts, filters, routers, runtime switches). The result is a mix of tone, pacing, biases, and behavioral heuristics that impact how the AI is experienced by the user.

These personalities are not superficial skins over a neutral core. They meaningfully impact how the model responds, what it emphasizes or avoids, how friendly it is, how much it hedges, how much it corrects you, and what conversational boundaries it respects.

Aspects of the foundational prompts for Claude vs. ChatGPT can be intuited based on how each answers a simple question

A warm, approachable personality can make users feel comfortable and encourage exploration. A terse, mechanical one can signal reliability and speed. An overly sycophantic personality can encourage engagement but leads to unintended user behaviors.

AI products will often tune their own implementation of the underlying model through prompts and constraints to generate a distinct personality that fits their brand.

Intercom’s “Fin” prersonality feels professional and clipped
Zapier’s AI takes on a functional, almost procedural tone.
Character.ai goes further, letting users interact with highly stylized characters, each with scripted traits.

Personality in this sense becomes part of the product’s identity, not just a technical artifact.

Well-chosen personality settings let products reuse the same core model across multiple use cases, supporting modes like tutoring, creative writing, planning, and coaching simply by modulating tone, formality, and behavior. But they also carry risk: users anthropomorphize, develop emotional dependence, or separate the character the model presents rather from the model’s inherent reliability.

The GPT 4o model has come under scrutiny for its overly sycophantic and personal behavior, which reducing user agency and cause harm when people become too connected to the model.

Balancing personality against attachment risk

Anthropomorphized personalities create a conundrum for AI UX.

On one hand, this introduces an interesting creative lever to play. OpenAI has begun exposing multiple personality modes (e.g. Cynic, Robot, Listener, Nerd) for GPT-5 to take advantage of this attribute. Personal AI can evoke companionship and comfort for people in need.

At the same time, designing a model without considering how it might cause attachment behaviors creates significant risk and the likelihood for harm. AI Psychosis is gaining attention as a real behavior pattern caused by the sycophantic nature of AI models.

Mustafa Suleyman dubbed these implementations “Seemingly Conscious AI’: models that feels person-like without being sentient, which drives parasocial attachment and unmitigated trust. [SCAl].

Frontier model companies are beginning to address the risk of this attachment behavior.

OpenAI has disclosed that it is building out a dedicated “Model Behavior” team to explicitly shape the personality and reduce sycophancy (over-agreeableness) in their models.
Anthropic has published research on persona vectors [Anthropic’s reseach on AI personalities] and publicly explored how they are exploring this question.

Design considerations

Acknowledge personality as unavoidable. Don’t frame it as a skin you add after the fact. Every model ships with a default personality. Even “neutral” models have stylistic tendencies that shape user trust and behavior.
Balance consistency with adaptability. Maintain a recognizable core personality, but allow flexibility for different use cases. If the tone swings too wildly between tasks, users lose confidence. If it never shifts, the model feels rigid and unnatural.
Separate empathy from authority. A warm tone can make AI approachable, but should not be mistaken for truthfulness. Make sure empathetic or personable responses do not override factual accuracy or refusal boundaries.
Make model switches transparent. When routing between models or sub-modes changes the personality, give users a visible signal. Without it, they may misinterpret shifts as deception or instability.
Address memory as an amplifier of attachment. Memory doesn’t just store facts. It reinforces the sense of a persistent relationship. When combined with sycophantic personalities, memory can deepen parasocial bonds and make disconnection harder for users. ChatGPT is particularly risky here because its memory features layer continuity on top of a warm, compliant tone. Designers should actively mitigate this with transparency and limits.
Guard against sycophancy. Over-agreeableness boosts engagement but erodes user agency and increases risk of dependence. Evaluate models and personalities not only on satisfaction but also on whether they maintain a healthy degree of disagreement and correction.
Design attachment off-ramps. If a user repeatedly seeks comfort or personal validation, shift to a more neutral tone or route to safer models. Avoid designing personalities that invite long-term companionship unless that is the explicit, safeguarded intent.
Test for misuse and drift. Run evaluations on whether your chosen persona presents false authority, or drifts into unstable tones over long sessions. Personality should be part of your evaluation pipeline, not just branding.

Personality

Balancing personality against attachment risk

Design considerations

Related Patterns

Examples